Created by W.Langdon from gp-bibliography.bib Revision:1.4910
In this dissertation we follow alternative, semantic-oriented approach that concentrates on the actual behavior of programs in population to determine how to construct the new ones. This research trend grew up as an attempt to overcome weaknesses of methods that rely only on syntax analysis. Recent contributions suggest that semantic extensions to GP can be a remedy to poor performance of classical, syntactic methods.
Therefore, in this dissertation we firstly present the advantages and disadvantages of several possible descriptions of program's behaviour. Then we introduce the concept of semantics used in all semantic extensions presented throughout this thesis.
The first semantic extension presented in this thesis is a method population initialisation which forces the individuals in population to be semantically unique. We also show selected semantic-aware variants of crossover and mutation operators. In particular, we test how they perform with and without our initialization method.
Next, we introduce and formalise our novel concept of desired semantics that describes the desired behaviour for given part of a program. Then we propose several methods that employ desired semantics to create new programs by combining matching parts. We show that some of these methods significantly outperform other methods, semantic as well as syntactic ones.
The second important proposition of this thesis is the concept of functional modularity. Functional modularity involves defining modules based on their semantic properties instead of syntactical ones, like, e.g., the frequency of occurring some code fragments. Functional modularity can be used to decompose a problem into potentially easier parts (subproblems), and then to solve the subproblems in isolation or together.
All the described methods are illustrated with extensive experimental verification of their performance on a carefully prepared benchmark suite that contains problems from various domains. On this suite, we show the overall advantage of semantic-aware extensions, especially for methods that rely on desired semantics.",
Genetic Programming entries for Bartosz Wieloch